What we learned building an AI that actually helps people get jobs (not just look like it helps)

We built UniqU (https://www.getuniqu.com) — an end-to-end AI job search platform. After 18 months in production, here are the things we got wrong first and what we learned. The ATS optimization problem is harder than it looks Every resume tool claims to 'beat ATS.' Most do keyword frequency counting. The problem is that modern applicant tracking systems — and the AI layers being added on top of them — increasingly evaluate keywords in context, not just presence. Our AI Resume Optimization had to move from 'does this keyword appear?' to 'does this keyword appear in a meaningful work context with a concrete outcome?' Building that semantic evaluation layer properly took three iterations. Voice coaching with real memory is a completely different product from a chatbot The AI Interview Coach was the most technically ambitious piece. For coaching to feel genuine, the model needs to: – Remember previous sessions – Know the user's resume deeply (not just have read it once) – Know the target role and company – Adapt to the session type (career coaching vs. interview simulation vs. feedback review) We built a memory architecture that persists across sessions and injects the right context dynamically. The difference between this and a stateless chatbot is the difference between a good coach and a random stranger answering career questions. Company research at interview depth requires more than web search Building a real AI Company Research Tool required going beyond surface-level search results. We needed to surface: strategic priorities from executive communications, culture signals from employee content patterns, recent organizational changes, and role-specific preparation angles. Synthesizing this into a pre-interview briefing that's actually useful (not just a wall of links) required significant prompt engineering and evaluation work. The product lesson Job search tools often optimize for feeling productive rather than being productive. We've tried to build the opposite: fewer features, higher impact per feature, measurable outcomes. Happy to answer questions on any of the technical architecture.

Comments

[Uniqu]

Really insightful breakdown. Most resume tools stop at keyword matching, but contextual evaluation is the real challenge. Platforms like UniqU pushing deeper AI-driven job search tools are exactly what candidates need today.